Content-based retrieval by multiple image examples for sign board retrieval

In the area of retrieving image databases, one of the promising approaches is to retrieve it by specifying image example. However, specifying a single image example is not always sufficient to get satisfactory result, since one image example does not give comprehensive ranges of values that reflect the various aspects of the object to be retrieved. In this paper, we propose a method of retrieving images by specifying multiple image examples that is designed for retrieving sign boards. Features of color, shape, and spatial relation of color regions are extracted from example images, and they are clustered so as to obtain proper range of values. Compared with QBE systems that accept only a single image as the query condition, MIERS (Multi-Image Example-based Retrieval System) returns better retrieval result, where the experimental result showed that specifying more examples helps to improve recall with little deterioration of precision.